Overview
llama.cpp 1.0 is a promise rather than a performance milestone. It shipped 1.0.0 alongside the ggml tensor library on Nov 13, and what actually changed is one sentence: from now on, breaking changes will be marked by the major version.
The rule is spelled out by change type. A breaking change to the public C API under include bumps the major version. A backward-compatible feature, a backend update, new model support or an added API bumps the minor. A bug fix with no API change bumps the patch. ggml went to 1.0 the same day and follows the same scheme. The team added that as adoption grows it will be stricter about making API and ABI boundaries visible to downstream consumers.
Why that matters more than a benchmark number. For three and a half years the project stayed on 0.x or simply tracked build numbers, leaving Python, Node, Rust and Go bindings to guess compatibility from dates and build ids, with every upgrade a small gamble on whether it still compiles. Semantic versioning lets a binding pin dependencies by major version and lets applications follow releases in CI. For anyone compiling inference into their own program, that is worth more than a few percent of throughput.
The underlying capability is unchanged, merely confirmed by 1.0: support for more than 150 architecture variants plus every mainstream speculative decoding method. Each backend compiles into a dynamic library that can be loaded at runtime, which is why the same code base travels from laptops to phones to development boards.
Eight distribution channels are listed, an easy column to skip but a practical one: prebuilt artifacts for Linux, macOS, Windows, Android and iOS, plus Docker, Debian, Homebrew, conda-forge, MacPorts and XCFramework, with Winget on Windows. Hardware support comes from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel, with self-hosted runners still being added.
Two things are missing and worth watching. No list of breaking changes shipped with the release, so what 1.0.0 changed against the previous build has to be read from release notes and diffs. And no roadmap appeared, with duplex or multimodal direction left unmentioned. Read 1.0 as the start of stability, not a feature boundary.
Key Features
- Semantic versioning for the C API: Only a breaking change to the public C API under include bumps the major version; compatible features and new model support bump the minor, and pure fixes bump the patch, so downstream can decide whether an upgrade is safe.
- ggml released 1.0 in step: The ggml tensor library shipped 1.0.0 the same day and follows the same rules, so the foundation and the inference layer make consistent promises.
- More than 150 architecture variants: Cumulative support covers over 150 model architecture variants plus every mainstream speculative decoding method, with new models usually runnable shortly after a community submission.
- Backends as runtime-loaded libraries: Each backend compiles into a dynamic library and is optionally loaded at runtime, so one program can move between CPUs, vendor GPUs and mobile chips without shipping a separate bundle.
- Eight distribution channels: Prebuilt artifacts cover Linux, macOS, Windows, Android and iOS, with Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and XCFramework alongside.
- Continuous integration across six vendors: NVIDIA, AMD, Qualcomm, Arm, IBM and Intel provide hardware support, and self-hosted runners keep growing so each platform is built and tested routinely.
Use Cases
- Embedding inference in desktop or mobile apps, relying on prebuilt packages and runtime backends to cut distribution work
- Running models offline or inside a network where data cannot leave
- Pinning a self-written binding to the major version and following minor releases safely in CI
- Reusing one inference code base across development boards and phones by loading the right backend
Pros
- The C API and ABI now carry an explicit stability promise, so downstream stops guessing compatibility from dates
- Runtime backend loading lets one program cover CPUs and several acceleration vendors
- Distribution is unusually complete, from package managers to XCFramework
- MIT licensed, with no fee for commercial use or modification
Pricing
MIT licensed and free, with no royalty for commercial use or redistribution. Costs sit in your own hardware and engineering. There is no hosted offering and no paid tier.
Summary
The value is not the number 1.0 but the version semantics now available to depend on. Bindings and applications can finally judge upgrade risk by major version, which is more useful to teams embedding inference than throughput figures. Watch that no breaking-change list or feature roadmap shipped with it; 1.0 marks the start of stability, nothing more.
Version History
- 1.0.0 release (2026-11-13): llama.cpp and the ggml tensor library shipped 1.0.0 together and adopted semantic versioning for the public C API: a breaking change under include bumps the major version, backward-compatible features, backend updates, model support and added APIs bump the minor, and fixes without API changes bump the patch. Cumulative support covers more than 150 architecture variants and all mainstream speculative decoding methods, with backends compiling into dynamic libraries loaded at runtime. Distribution spans prebuilt artifacts for Linux, macOS, Windows, Android and iOS plus Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and XCFramework. Hardware support comes from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel. No breaking-change list was published with the release.